A wearable ultrasound patch can image the heart continuously. Running such a patch on a battery means it cannot stream raw video to a server, and it cannot run a large segmentation model on the device either. It needs a rule for which frames are worth the radio energy of sending them off for careful analysis. Such a rule needs a notion of how wrong the on-device estimate is allowed to be.
Cardiology already supplies one. Expert sonographers tracing the same left ventricle disagree by a measurable and frame-dependent amount. In ejection fraction, the fraction of blood pumped per beat, that disagreement is reported at 7–13%. We propose to make this human disagreement the tolerance for wireless offloading. The device sends a frame for heavy server-side analysis only when its own estimate is likely to differ from the server's by more than experts would differ from each other on that same frame. Otherwise it sends a few bytes and the radio stays asleep.
The design rests on an observation that runs against the obvious version of this idea. When a frame is ambiguous because experts genuinely disagree about where the boundary is, sending it to a bigger model buys nothing, because the bigger model cannot settle a question that the anatomy and image do not settle. Radio energy should go to frames where the small model is uncertain, not to frames where the experts are.
Core Claim
On a battery-limited cardiac imager, the uncertainty that should trigger offloading is the on-device model's expected error relative to a strong server model, scaled by a per-frame estimate of expert spread. Raw uncertainty, which mixes model error with expert disagreement, is the wrong trigger. Concretely, let $\hat e_t$ be the device's predicted gap between its own and the server's left-ventricle area at frame $t$. Let $\hat s_t$ be the device's predicted standard deviation of that area across expert tracers. The device transmits the full frame (or a cropped region around the ventricle) when
$$\hat e_t > \lambda\, \hat s_t ,$$
where $\lambda$ is one knob trading energy for fidelity. Every other frame, it sends a compact packet containing the area estimate and $\hat s_t$.
We predict three things. First, at matched energy this rule keeps more frames within inter-observer tolerance than fixed-interval transmission, change-detection filters, or plain uncertainty thresholds. Second, the gain comes specifically from dividing by a per-frame $\hat s_t$ rather than a constant. Third, the gain grows as image quality drops, because low-quality frames inflate expert disagreement and fool raw-uncertainty triggers into wasting transmissions.
What Exists and What Is Missing
The application regime comes from A wearable cardiac ultrasound imager. It built a stretchable patch with an orthogonal transducer array and fed its stream to a deterministic segmentation network (FCN-32) that tracks left-ventricular volume through rest and exercise. The patch is tethered to a workstation. It has no wireless budget, no duty cycling, and no notion of rater ambiguity. Our work supplies exactly the piece it leaves open: deciding, frame by frame, what a wireless version should spend energy on.
The gating architecture borrows from sensing systems that place a cheap always-on sentinel in front of an expensive stage. DSP.Ear runs silence and speech admission filters on a phone's low-power co-processor, reuses prior labels when features are stable, and escalates to the main processor only for rare complex events. It reports 3–7x efficiency gains over running the same work on the main processor. Somniloquy lets a small embedded processor stand in for a sleeping PC, with 11–24x lower power than an idle PC. Both supply the method: a two-tier design with a cheap gate. Both gate on signal statistics, not on whether the escalation would change a clinically meaningful answer, and that is the part we replace.
The disagreement modelling comes from two papers. A Probabilistic U-Net for Segmentation of Ambiguous Images attaches a low-dimensional latent variable to a U-Net, a standard encoder–decoder segmentation network. It injects samples only at the last layers, so many plausible contours cost little more than one forward pass. Both properties matter on a microcontroller. Pionono gives each annotator a Gaussian in latent space, whose mean captures that rater's habitual bias and whose variance captures their inconsistency. It reports Cohen's κ of 0.758 on Gleason 2019. We reuse both: late injection to keep sampling cheap on-device, and rater-specific latents to produce $\hat s_t$ in units of real expert spread. Both papers ran on server GPUs on CT and pathology images, and neither uses its uncertainty to drive any system decision.
Several papers bound what the server and the device can do. Foundation versus Domain-Specific Models for Left Ventricular Segmentation on Cardiac Ultrasound shows that a fine-tuned Segment Anything Model (SAM), a large general-purpose segmentation model, is far more robust on handheld point-of-care scans than a task-specific network. It reports Dice overlap of 0.857 against 0.667. This makes SAM the natural server model. Its vision-transformer backbone also cannot run on a patch, which is why the device needs a small model and a gate. SimLVSeg reaches 93.32% Dice on EchoNet-Dynamic using labels on fewer than 1.2% of frames, with 3.8x fewer operations than sliding-window methods. It shows a lighter video model is possible, but it counts only arithmetic and ignores radio energy and rater spread. Fully automated quantification of cardiac chamber and function assessment in 2-D echocardiography shows two things: automated accuracy falls on poor-quality scans, and inter-observer agreement on the same quantities ranges from ICC 0.899 to 0.982. That variation across conditions is the motivation for a per-frame rather than a fixed tolerance.
Finally, Calibrate the inter-observer segmentation uncertainty via diagnosis-first principle argues that rater labels should be weighted by their usefulness to a downstream diagnosis. That partly cuts against us, since we treat the spread of rater opinion as the tolerance rather than collapsing it. We do not adopt diagnosis-first weighting, because it ties the target to one particular downstream classifier, a limitation that paper itself notes. We do test it as an alternative way to define the reference.
Research Program
The device-side model is a small 2D encoder–decoder sized to fit microcontroller memory, quantized to 8-bit. It has two heads. The first is a Pionono-style rater latent injected late, so drawing K contours is cheap. The spread of left-ventricle area across those draws, mapped through rater-specific latents, gives $\hat s_t$. The second is a scalar regressor trained to predict $\hat e_t$, the gap between the device's area and the server's area. It is supervised by running the server model offline on the training set.
The server model is SAM fine-tuned as in the foundation-model comparison paper above, with a probabilistic head trained on the same multi-rater labels. The server therefore also returns an expert-spread estimate, which lets us check that $\hat s_t$ is calibrated.
Data is the main practical obstacle. EchoNet-Dynamic and CAMUS are described in the file only with single consensus traces. We will therefore collect a multi-rater subset: three to four trained sonographers re-tracing end-diastolic and end-systolic frames for several hundred EchoNet-Dynamic videos, plus a quality-stratified sample. This is the tracing workload that the foundation-model paper measured at roughly 6–12 seconds per frame, so it is feasible.
LIDC-IDRI's four-radiologist labels serve a narrower role. We use them only to check that the rater-latent decomposition separates expert spread from model error, not for the cardiac claims.
Energy accounting uses a real low-power board with a microcontroller and a low-power radio. We measure per-inference compute energy, per-byte transmit energy, and wake-up overhead, then replay every gating policy over the same test videos. Every policy is charged with the same measured costs.
We will also degrade the input to imitate a cheaper acquisition, since the wearable imager reports needing wide-beam compounding to get adequate image quality. The degradations are fewer compounded angles approximated by added speckle and contrast loss, plus lower frame rate. This lets us measure how acquisition savings inflate expert spread, an effect no paper in the file quantifies.
Baselines
The naive reference points are always-transmit and never-transmit. In always-transmit, every frame goes to the server SAM. This gives the fidelity ceiling at maximal radio cost; its reported EchoNet-Dynamic Dice is 0.911, against 0.915 for the EchoNet model, per the foundation-model comparison paper. In never-transmit, only the on-device model is used, giving the energy floor. SimLVSeg's 93.32% Dice on EchoNet-Dynamic stands as the strongest reported single-consensus accuracy for comparison, but it is not a gating policy.
Fixed-interval transmission sends every N-th frame plus the compact packet otherwise, sweeping N to trace an energy–fidelity curve.
Signal-change admission filtering, modelled on DSP.Ear, transmits when frame-to-frame pixel or feature change exceeds a threshold. It is the strongest systems competitor because it needs no labels and costs almost nothing to compute.
Raw-uncertainty gating is the obvious intersection idea. It transmits when the total spread of the device's sampled contours is high, with no split into model error and expert spread.
We also run two learned gates. Learned deferral without a tolerance thresholds $\hat e_t$ against a constant; it is the main threat. Diagnosis-first variant defines the reference contour with weights in the style of Calibrate the inter-observer segmentation uncertainty via diagnosis-first principle instead of using the rater distribution.
Two ablations isolate the key ingredient: replacing $\hat s_t$ with its dataset average, and replacing it with the server-computed spread, an oracle. Every policy sees the same videos, the same measured energy costs, and the same on-device network weights. Each policy's single threshold is tuned on the same validation split, and policies are compared at matched mean power.
The proposal beats the field if it does better on the full energy-versus-fidelity curve. Fidelity here is the fraction of frames whose area error relative to the server lies within one expert standard deviation, plus ejection-fraction error per beat. On that curve, the proposal must sit significantly above learned deferral and signal-change filtering across most of the power range. It must also close a large share of the gap to the oracle-tolerance ablation.
Why not just…?
Why not just transmit when the model is uncertain? Because on ambiguous frames the server inherits the same ambiguity, so sending them buys nothing. The raw-uncertainty baseline tests this directly. The prediction is that it spends a disproportionate share of transmissions on low-quality frames where expert spread is wide. We will report that share as a diagnostic. If raw uncertainty matches our rule, the expert-disagreement component is not doing work and the claim fails.
Why not just send every N-th frame? Cardiac events of interest are transient. DSP.Ear built content-aware filters precisely because blind periodic sampling misses short events. Still, for a steady resting heart, periodic sampling may be nearly optimal. We will report results separately for stable segments and for segments with abrupt ejection-fraction or rhythm changes, and concede the stable regime if the curves coincide there.
Why not just gate on frame-to-frame change? Change and error are different things. A beating ventricle changes every frame, yet a small model may track it well. A nearly static but poorly imaged wall may be badly wrong. This filter is a real threat, because it is nearly free. The matched-energy comparison settles it, and a hybrid of change detection with our tolerance is included in case the two signals turn out to be complementary.
Why not just train a deferral predictor and threshold it with a constant? This is the closest competitor, and it may capture most of the benefit. The constant-tolerance ablation is designed to show whether per-frame expert spread adds anything. Large quality variation in the data is what should make a per-frame tolerance matter, given ICC from 0.899 to 0.982 in Fully automated quantification of cardiac chamber and function assessment in 2-D echocardiography and the drop on handheld scans in Foundation versus Domain-Specific Models for Left Ventricular Segmentation on Cardiac Ultrasound. If the curves coincide even on the degraded-acquisition arm, we will report that the rater model is unnecessary for gating and keep only its calibration findings.
Proposed Experiments and Decisive Checks
The primary experiment replays all policies over held-out EchoNet-Dynamic videos and the multi-rater subset. It reports measured mean power against two fidelity measures: the within-tolerance fraction, and ejection-fraction mean absolute error against the multi-rater mean. The foundation-model comparison paper reports 5–7% ejection-fraction error for its models against 7–13% expert variation, which gives the scale for judging whether a policy stays clinically acceptable. CAMUS serves as a domain-shift test without retraining.
The decisive check is the constant-versus-per-frame tolerance ablation, run across the degraded-acquisition arm. If our rule's advantage does not widen as image quality falls, the mechanism is wrong. A second check tests calibration of $\hat s_t$: the device's predicted expert spread should correlate with observed spread among the held-out sonographers, frame by frame. Without that correlation, the method is only deferral under another name.
A third analysis maps how reduced acquisition effort, such as fewer compounded angles or a lower frame rate, increases expert spread and changes the optimal transmission rate. This gives the first measured link between transmit energy and inter-observer stability for a wearable imager.
The main risk is that radio cost on real hardware dwarfs every compute difference, making the gate's choices less important than packet size. Reporting component-wise energy will show whether that is the case.
References
- A wearable cardiac ultrasound imager
- DSP.Ear: leveraging co-processor support for continuous audio sensing on smartphones
- Somniloquy: Augmenting Network Interfaces to Reduce PC Energy Usage
- A Probabilistic U-Net for Segmentation of Ambiguous Images
- Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation
- Foundation versus Domain-Specific Models for Left Ventricular Segmentation on Cardiac Ultrasound
- SimLVSeg: Simplifying Left Ventricular Segmentation in 2D+Time Echocardiograms with Self- and Weakly-Supervised Learning
- Fully automated quantification of cardiac chamber and function assessment in 2-D echocardiography: clinical feasibility of deep learning-based algorithms
- Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle